- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper examines the relationship between regular and irregular time series forecasting. It organizes representative methods into four stages—normalization, input embedding, dependency modeling, and forecast generation—and develops BridgeTST through comparisons at each stage. The model combines observed values, masks, timestamps, and variable identities on a shared timeline, using general-purpose embedding, attention, and forecasting components. Experiments cover both regular and irregular data, with module replacements used to examine the role of specialized designs.
Critical strengths
- The research question is valuable. The paper distinguishes backbone capabilities from the roles of input information and the processing pipeline. This helps reassess why models designed for regular data underperform on irregular data and offers a useful direction for building stronger baselines.
- The analysis has a reasonably clear empirical basis. The four-stage framework clarifies the common structure of different methods. Comparisons of normalization, input information, cross-variable interactions, and prediction heads support the selected design choices.
- The method has practical value in both settings. BridgeTST is competitive on the evaluated regular and irregular datasets, supporting its architecture built from general-purpose components as an important reference for irregular forecasting.
Critical weaknesses
- The handling of future target timestamps needs clarification. The appendix specifies each dataset's alignment resolution and forecast horizon, but does not make clear whether these settings allow target timestamps to be inferred from output positions. Equation (3) conditions forecasts on future query timestamps, whereas Equation (8) and the architecture diagram do not show how they are used. Clarifying this correspondence would establish which query settings the fixed-length prediction head supports.
- Direct evidence for why existing regular forecasting models underperform remains limited. The stage-by-stage analysis supports BridgeTST's effectiveness, but the paper further suggests that the limitations of regular forecasting models may stem from normalization and input information rather than their backbones. The main experiments do not sufficiently explain how the regular baselines were adapted in these respects, or clearly demonstrate the effects of such adaptations on representative existing backbones. The current evidence more directly supports the claim that general-purpose components can form an effective model; it remains insufficient to establish the main causes of the performance gaps of existing regular models.
Questions for authors
- Do the alignment and forecasting settings in the appendix mean that the future target timestamps in Equation (3) are implicitly determined by output positions and need not be supplied explicitly? Please clarify how this relates to Equation (8). If target timestamps are not fixed, how are outputs matched to query timestamps?
- The paper suggests that the limitations of existing regular forecasting models on irregular data may arise from normalization and the handling of sampling information. Do the experiments support this explanation, or only establish that these design choices are effective within BridgeTST? Please clarify the scope of the conclusion accordingly.
- Figure 3 shows that the module replacements degrade performance, but does not establish the specific causes of that degradation. The proposed mechanistic explanation should therefore be qualified.
4 comments
Response to Reviewer Kipt (Part II)
W2&Q2 (continued: input-feature ablation).
| Model and setting | HumanActivity | USHCN | P12 | MIMIC-III | MIMIC-IV |
|---|---|---|---|---|---|
| Metric | MSE / MAE | MSE / MAE | MSE / MAE | MSE / MAE | MSE / MAE |
| PatchTST: no RevIN, values only | 0.0606 / 0.1640 | 0.1558 / 0.2453 | 0.3277 / 0.3720 | 0.4824 / 0.4176 | 0.3574 / 0.3945 |
| PatchTST: no RevIN, | 0.0547 / 0.1510 | 0.1775 / 0.2868 | 0.3169 / 0.3649 | 0.4648 / 0.4057 | 0.2891 / 0.3291 |
| PatchTST: no RevIN, | 0.0667 / 0.1809 | 0.1579 / 0.2345 | 0.3161 / 0.3625 | 0.4479 / 0.3923 | 0.2678 / 0.3098 |
| PatchTST: no RevIN, | 0.0668 / 0.1762 | 0.1543 / 0.2302 | 0.3157 / 0.3613 | 0.4408 / 0.3868 | 0.2611 / 0.3035 |
Overall, these results lead to conclusions broadly consistent with those drawn from BridgeTST, with some differences across individual datasets. Special design choices in regular models may introduce additional factors. For example, PatchTST processes each channel independently. This may lead to much worse performance on irregular data than irregular-specific models and obscure the underlying cause. BridgeTST starts from a standard patch-based architecture commonly used for regular forecasting. This makes it a more controlled starting point than regular models with diverse architectures.
The PatchTST normalization and input ablations are reported in .
Q3: The module-replacement results show a performance drop but do not identify its cause. The paper should therefore limit its mechanistic explanation.
A: To better understand the performance degradation observed in Figure 3, we further conduct ablation studies on TTCN.
TTCN forms a patch representation by weighting each position's input with a position-specific filter coefficient, applying the same operation across output channels. For one output channel, let index a patch position; its input combines the observed value and timestamp encoding:
Here, is the observed value, indicates whether the position is valid, and is the timestamp encoding. The original TTCN computes coefficients from the current input. We also introduce a static ablation in which the coefficients do not depend on the input. They are learned directly:
g_\theta(\mathbf u_t), & \text{TTCN},\\\\ \mathbf s_t, & \text{static TTCN}. \end{cases}$$ We also study how TTCN aggregates observations using these coefficients. The original TTCN uses softmax. We compare it with raw sum and mean sum, where $\mathcal V$ denotes the valid time positions: $$\mathbf w_t=\begin{cases} \operatorname{softmax}_{t\in\mathcal V}(\mathbf a_t), & \text{softmax},\\\\ \mathbf a_t, & \text{raw sum},\\\\ \mathbf a_t/\max(1,|\mathcal V|), & \text{mean sum}. \end{cases}$$ Softmax normalizes coefficients over valid positions. Raw sum uses the coefficients directly. Mean sum divides the sum by the number of valid positions. The weighted inputs are then summed to produce the output for that channel: $$z=\sum_{t\in\mathcal V}\mathbf w_t^\top\mathbf u_t.$$ We replace the input embedding in BridgeTST and compare five settings on ETTh1 and ETTm1 with an input length of 512 and a forecast horizon of 96: MLP, Original TTCN, Static TTCN with softmax, Static TTCN with raw sum, and Static TTCN with mean sum. | Input embedding | ETTh1 | ETTm1 | |---|---:|---:| | **Metric** | **MSE / MAE** | **MSE / MAE** | | MLP | **0.3650 / 0.3892** | **0.2818 / 0.3281** | | Original TTCN | 0.6830 / 0.5550 | 0.3745 / 0.4135 | | Static TTCN / softmax | 0.6504 / 0.5737 | 0.3332 / 0.3811 | | Static / raw sum | 0.4441 / 0.4585 | 0.3040 / 0.3488 | | Static / mean sum | 0.5974 / 0.5441 | 0.3442 / 0.3837 | Replacing softmax with raw sum substantially reduces the error, but the result is still worse than MLP. This suggests that softmax aggregation accounts for part of TTCN's performance drop. We have limited the mechanistic claims in the revised manuscript accordingly. The TTCN ablation is described in $\underline{\text{Section 5.3.2}}$, with results in $\underline{\text{Table 7}}$.Official Comment by Reviewer Kipt
Thank you for the clarifications and additional experiments, which substantially address my main concerns about the forecasting setup and the evidence supporting the paper’s conclusions. I have no further questions and keep my score.
Thank You for Your Feedback
Thank you for reviewing our clarifications and additional experiments. We appreciate your feedback and are glad that our response addressed your concerns.
Response to Reviewer Kipt (Part I)
Thank you for your detailed comments on unclear equations, evidence for adapting regular models, and the explanation of module-replacement effects.
W1&Q1: The use of future target timestamps needs clarification.
A:
Equation (3) uses the set of future target timestamps Q as a prediction condition. The MLP head in Equation (9) always outputs F ordered predictions and does not explicitly read Q. Under our fixed-window forecasting protocol, the f-th output corresponds to the f-th target time step. Thus, the head supports a fixed, ordered forecast window. It does not support queries at arbitrary timestamps.
We compare the FTC Head with and without future timestamps. The results are shown below.
Across the five irregular datasets, future timestamps have only a small effect on performance. In the settings studied here, explicitly providing future timestamps makes little difference.
W2&Q2: Direct evidence about why regular forecasting models perform poorly is limited.
A: To further study this performance gap, we conduct adaptation experiments on PatchTST.